Quite a few small models, not many large models (which are not gigantic frontier models either for the record).

The datasets used are all public domain or expired copyrights, like Common Corpus and Common Pile. There are also models like Talkie that are exclusively based on data from prior to 1930.

So you've got the Pleias Common Models, Comma (EleutherAI and people on Huggingface contributing), KL3M (Kelvin Legal Large Language Model), Apertus (ETH Zurich), and OLMo/v2 (Open Language Model, Allen Institute for AI) as good prime examples.

@Osm@lemmy.ml for your reference

Not a clue myself, but when I have to choose, I go with Mistral. They used to have a very good policy last time I've read it. They're also targeting the European market a lot (and are located in Europe), which seems much safer than US or China in terms of data handling to me.

They also shared open-weight state-of-the-art models and ideas multiple times. The 7B Mistral was a thing for a while, and they were the first to ever publish an open-weight MoE model (Mixtral) I believe.

You're absolutely right no AI model is completely free of ethical flaws when it comes to training data.

But there’s a fundamental difference between models that have already been trained and those that are currently being trained using my data and files. I’m looking for the former—tools that respect privacy and don’t use my conversations to improve their systems (most people recommend thaura.ai for this purpose).

Furthermore, I’m looking for options that are less involved in ongoing human rights violations—no military contracts, no ties to governments involved in occupation, and so on.